dearwolf/aider
1
0
Fork 0
mirror of https://github.com/Aider-AI/aider.git synced 2025-05-31 09:44:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 9

copy

Browse source
This commit is contained in:
Paul Gauthier 2023-07-02 10:39:50 -07:00
parent 4ea70bfdc0
commit fc519ca6b8
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
docs/benchmarks.md
View file

@ -35,10 +35,10 @@ I ran the benchmark
on all the ChatGPT models (except `gpt-4-32k`), using a variety of edit formats.
The results were interesting:
- Asking GPT to return an updated copy of the whole file in a standard markdown fenced code block proved to be the most reliable and effective edit format across all GPT-3.5 and GPT-4 models. The results for this `whole` edit format are shown in solid blue in the graph.
- Using the new functions API for edits performed worse than the above whole file method, for all the models. GPT-3.5 especially produced inferior code and frequently mangled this output format. This was surprising, as the functions API was introduced to enhance the reliability of structured outputs. The results for these `...-func` edit methods are shown as patterned bars in the graph (both green and blue).
- The performance of the new June (`0613`) versions of GPT-3.5 appears to be a bit worse than the February (`0301`) version. This is visible if you look at the "first attempt" markers on the first three solid blue bars and also by comparing the first three solid green `diff` bars.
- As expected, the GPT-4 models outperformed the GPT-3.5 models in code editing.
- **Plain text edit formats worked best.** Asking GPT to return an updated copy of the whole file in a standard markdown fenced code block proved to be the most reliable and effective edit format across all GPT-3.5 and GPT-4 models. The results for this `whole` edit format are shown in solid blue in the graph.
- **Function calls performed worse.** Using the new functions API for edits performed worse than the above whole file method, for all the models. GPT-3.5 especially produced inferior code and frequently mangled this output format. This was surprising, as the functions API was introduced to enhance the reliability of structured outputs. The results for these `...-func` edit methods are shown as patterned bars in the graph (both green and blue).
- **The new GPT-3.5 models did worse than the old model.** The performance of the new June (`0613`) versions of GPT-3.5 appears to be a bit worse than the February (`0301`) version. This is visible if you look at the "first attempt" markers on the first three solid blue bars and also by comparing the first three solid green `diff` bars.
- **GPT-4 does better than GPT-3.5,** as expected.
The quantitative benchmark results agree with my intuitions
about prompting GPT for complex tasks like coding. It's beneficial to